On 5/9/2019 8:01 PM, Benjamin Herrenschmidt wrote:
On Thu, 2019-05-09 at 16:51 -0700, Jae Hyun Yoo wrote:
On 5/9/2019 4:19 PM, Benjamin Herrenschmidt wrote:
On Thu, 2019-05-09 at 10:19 -0700, Jae Hyun Yoo wrote:
Okay. Probably I need to add one another patch in this series to
address
what you pointed out.
I have one question. I reviewed again all bitops in this driver
and
checked that some bitops are protected by a spinlock and some
others
are not. In this case, can I mix use atomic and non-atomic bitops
depend on each bitop's protection state by the spinlock? Or,
would it be
better to change all of them to bool in this case?
No, if some aren't protected by a lock and some are, then they need
to
remain atomic.
The question then becomes whether the unprotected ones are actually
correct or just exposing more races.
Got it. Not sure yet but I think the protected bitops could be moved
out
from the spinlock scope then all bitops could be kept as atomic.
Which is very likely to be extremely racy... (and gratuitously more
costly) :-)
Okay then, will try to wrap all bitops using a single spinlock and
change all of them to non-atomic bitops. Is it right approach?
Thanks,
Jae
I'll
look at and test this driver code more deeply again, and will submit
v2
soon.
Again, thanks a lot for your review.
Regards,
Jae
